Lack of stability guarantees strongly limits the use of reinforcement learning (RL) in safety critical robotic applications. Here we propose a control system architecture for continuous RL control and derive corresponding stability theorems via contraction analysis, yielding constraints on the network weights to ensure stability. The control architecture can be implemented in general RL algorithms and improve their stability, robustness, and sample efficiency. We demonstrate the importance and benefits of such guarantees for RL on two standard examples, PPO learning of a 2D problem and HIRO learning of maze tasks.
View on arXiv